oa::FnMatrix

FnMatrix namespace in the OA Core public surface.

NsC++ public headers

Public Functions

oa::Matrix oa::FnMatrix::abs(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::add(const oa::Matrix & inA, const oa::Matrix & inB)

void oa::FnMatrix::addInPlace(Matrix & inSelf, const Matrix & inOther)

oa::Matrix oa::FnMatrix::addScalar(const oa::Matrix & inA, oa::F32 inScalar)

void oa::FnMatrix::addScalarInPlace(Matrix & inSelf, oa::F32 inScalar)

oa::I64 oa::FnMatrix::argmax(const oa::Matrix & inA, oa::I32 inDim = -1)

Matrix oa::FnMatrix::avgPool2d(const Matrix & inX, oa::I32 inKernelSize, oa::I32 inStride, oa::I32 inPadding)

Matrix oa::FnMatrix::avgPool2dBwd(const Matrix & inX, const Matrix & inGradOutput, oa::I32 inKernelSize, oa::I32 inStride, oa::I32 inPadding)

BatchNorm2dBwdResult oa::FnMatrix::batchNorm2dBwd(const Matrix & inX, const Matrix & inGamma, const Matrix & inBeta, const Matrix & inMean, const Matrix & inVar, const Matrix & inOut, const Matrix & inDOut, oa::F32 inEps, bool inIsTraining)

oa::Matrix oa::FnMatrix::biasAdd(const oa::Matrix & inA, const oa::Matrix & inB)

Matrix oa::FnMatrix::bmm(const Matrix & inA, const Matrix & inB)

Matrix oa::FnMatrix::bmmNt(const Matrix & inA, const Matrix & inB)

Matrix oa::FnMatrix::cast(const Matrix & inSrc, oa::ScalarType inDtype)

void oa::FnMatrix::castInto(const Matrix & inSrc, Matrix & outDst)

Matrix oa::FnMatrix::categoricalAccuracyCount(const Matrix & inLogits, const Matrix & inLabels)

Matrix oa::FnMatrix::causalMask(oa::I64 inSeqLen)

Matrix oa::FnMatrix::causalMaskBwd(const Matrix & inGradOut)

Matrix oa::FnMatrix::channelNorm(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::I32 inBatch, oa::I32 inChannels, oa::I32 inSeqLen, oa::F32 inEps)

ChannelNormBwdResult oa::FnMatrix::channelNormBwd(const Matrix & inX, const Matrix & inWeight, const Matrix & inGradOutput, oa::I32 inBatch, oa::I32 inChannels, oa::I32 inSeqLen, oa::F32 inEps)

Matrix oa::FnMatrix::channelNormRelu(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::I32 inBatch, oa::I32 inChannels, oa::I32 inSeqLen, oa::F32 inEps)

ChannelNormBwdResult oa::FnMatrix::channelNormReluBwd(const Matrix & inX, const Matrix & inWeight, const Matrix & inFwdOut, const Matrix & inGradOutput, oa::I32 inBatch, oa::I32 inChannels, oa::I32 inSeqLen, oa::F32 inEps)

oa::Matrix oa::FnMatrix::clampMax(const oa::Matrix & inA, oa::F32 inMax)

oa::Matrix oa::FnMatrix::clampMin(const oa::Matrix & inA, oa::F32 inMin)

Matrix oa::FnMatrix::col2Im1d(const Matrix & inDCols, oa::I32 inN, oa::I32 inC, oa::I32 inL, oa::I32 inK, oa::I32 inStride, oa::I32 inPadding, oa::I32 inDilation, oa::I32 inOutL)

CompactRowsResult oa::FnMatrix::compactRows(const Matrix & inSelf, const Matrix & inMask)

Matrix oa::FnMatrix::compactRowsBwd(const Matrix & inGradOut, const Matrix & inRowMap, const Matrix & inCount, MatrixShape inInputShape)

Matrix oa::FnMatrix::concat(oa::Span<Matrix> inInputs, oa::I32 inDim = 0)

Matrix oa::FnMatrix::conv1dBwdData(const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding, oa::U32 inDilation, const MatrixShape & inInputShape)

Conv1dBwdWeightResult oa::FnMatrix::conv1dBwdWeight(const Matrix & inInput, const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding, oa::U32 inDilation)

Matrix oa::FnMatrix::conv1dGemm(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::I32 inStride = 1, oa::I32 inPadding = 0, oa::I32 inDilation = 1)

Matrix oa::FnMatrix::conv1dReluGemm(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::I32 inStride = 1, oa::I32 inPadding = 0, oa::I32 inDilation = 1)

Matrix oa::FnMatrix::conv2d(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::U32 inStride, oa::U32 inPadding, oa::U32 inGroups = 1)

Matrix oa::FnMatrix::conv2dBwdData(const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding, const MatrixShape & inInputShape, oa::U32 inGroups = 1)

Conv2dBwdWeightResult oa::FnMatrix::conv2dBwdWeight(const Matrix & inInput, const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding, oa::U32 inGroups = 1)

Matrix oa::FnMatrix::convTranspose2d(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, oa::U32 inStride, oa::U32 inPadding)

Matrix oa::FnMatrix::convTranspose2dBwdData(const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding, const MatrixShape & inInputShape)

ConvTranspose2dBwdWeightResult oa::FnMatrix::convTranspose2dBwdWeight(const Matrix & inInput, const Matrix & inDOut, const Matrix & inWeight, oa::U32 inStride, oa::U32 inPadding)

oa::Matrix oa::FnMatrix::copy(const oa::Matrix & inA)

oa::Status oa::FnMatrix::copyToHost(const Matrix & inSrc, void * outHost, oa::U64 inBytes)

oa::Matrix oa::FnMatrix::cos(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::dequantize(const oa::QuantMatrix & inInput)

Matrix oa::FnMatrix::detach(const Matrix & inSelf)

oa::Matrix oa::FnMatrix::div(const oa::Matrix & inA, const oa::Matrix & inB)

void oa::FnMatrix::divInPlace(Matrix & inSelf, const Matrix & inOther)

oa::Matrix oa::FnMatrix::divScalar(const oa::Matrix & inA, oa::F32 inScalar)

void oa::FnMatrix::divScalarInPlace(Matrix & inSelf, oa::F32 inScalar)

Matrix oa::FnMatrix::dropout(const Matrix & inA, oa::F32 inP, oa::U64 inSeed = 0)

oa::Matrix oa::FnMatrix::elu(const oa::Matrix & inA, oa::F32 inAlpha)

Matrix oa::FnMatrix::eluBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput, oa::F32 inAlpha = 1.0f)

Matrix oa::FnMatrix::empty(MatrixShape inShape, oa::ScalarType inDtype = weightDtype(), oa::MemoryPlacement inPlacement = oa::MemoryPlacement::Auto)

Matrix oa::FnMatrix::empyrealmAdt(const Matrix & inDdA, const Matrix & inDt, oa::F32 inAFloor)

EmpyrealmAdtBwdResult oa::FnMatrix::empyrealmAdtBwd(const Matrix & inDOut, const Matrix & inDdA, const Matrix & inDt, oa::F32 inAFloor)

Matrix oa::FnMatrix::empyrealmDt(const Matrix & inX, oa::F32 inDtMin, oa::F32 inDtMax)

EmpyrealmDtAdtResult oa::FnMatrix::empyrealmDtAdt(const Matrix & inDtRaw, const Matrix & inDdA, oa::F32 inDtMin, oa::F32 inDtMax, oa::F32 inAFloor)

Matrix oa::FnMatrix::empyrealmDtBwd(const Matrix & inDOut, const Matrix & inX, oa::F32 inDtMin, oa::F32 inDtMax)

Mamba3PreprocessResult oa::FnMatrix::empyrealmPreprocess(const Matrix & inProjected, const Matrix & inDtBias, const Mamba3PreprocessConfig & inConfig)

Mamba3PreprocessBwdResult oa::FnMatrix::empyrealmPreprocessBwd(const Matrix & inProjected, const Matrix & inDtBias, const Matrix & inDZ, const Matrix & inDX, const Matrix & inDBh, const Matrix & inDCh, const Matrix & inDDT, const Matrix & inDADT, const Matrix & inDTrap, const Matrix & inDAngle, const Mamba3PreprocessConfig & inConfig)

Matrix oa::FnMatrix::empyrealmSiso(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const SsmConfig & inConfig)

SsmBwdResult oa::FnMatrix::empyrealmSisoBwd(const Matrix & inDOut, const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const SsmConfig & inConfig)

Matrix oa::FnMatrix::empyrealmSisoStep(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const Matrix & inSsmState, const Matrix & inAngleState, const Matrix & inKState, const Matrix & inVState, const SsmConfig & inConfig)

Matrix oa::FnMatrix::equal(const Matrix & inA, oa::F32 inValue)

oa::Matrix oa::FnMatrix::exp(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::fill(const oa::MatrixShape & inShape, oa::F32 inValue)

void oa::FnMatrix::fillInPlace(oa::Matrix & inSelf, oa::F32 inValue)

Matrix oa::FnMatrix::flashAttentionCausal(const Matrix & inQ, const Matrix & inK, const Matrix & inV, oa::F32 inScale)

FlashAttentionBwdResult oa::FnMatrix::flashAttentionCausalBwd(const Matrix & inQ, const Matrix & inK, const Matrix & inV, const Matrix & inOutput, const Matrix & inLogSumExp, const Matrix & inGradOutput, oa::F32 inScale)

Matrix oa::FnMatrix::fromBytes(oa::Span<const oa::U8> inData, MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::fromInt32(oa::Span<const oa::I32> inData, MatrixShape inShape, oa::ScalarType inDtype = oa::ScalarType::Int32)

Matrix oa::FnMatrix::full(MatrixShape inShape, oa::F64 inValue, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::gather(const Matrix & inSelf, const Matrix & inIndices)

Matrix oa::FnMatrix::gatherBwd(const Matrix & inIndices, const Matrix & inGradOutput, oa::I32 inVocabSize, oa::I32 inEmbedDim)

Matrix oa::FnMatrix::gatherLastDim(const Matrix & inSelf, const Matrix & inIndices)

Matrix oa::FnMatrix::gatherLastDimBwd(const Matrix & inGradOut, const Matrix & inIndices, oa::I32 inInputWidth)

oa::Matrix oa::FnMatrix::geglu(const oa::Matrix & inA, oa::U32 inIntermediateSize)

Matrix oa::FnMatrix::gegluBwd(const Matrix & inInput, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::gelu(const oa::Matrix & inA)

Matrix oa::FnMatrix::geluBwd(const Matrix & inInput, const Matrix & inGradOutput)

Matrix oa::FnMatrix::greaterEqual(const Matrix & inA, oa::F32 inValue)

Matrix oa::FnMatrix::groupedGemmM(const Matrix & inX, const Matrix & inWeight, const Matrix & inOffsets)

GroupedGemmMBwdResult oa::FnMatrix::groupedGemmMBwd(const Matrix & inDOut, const Matrix & inX, const Matrix & inWeight, const Matrix & inOffsets)

Matrix oa::FnMatrix::groupedLinearM(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, const Matrix & inOffsets)

Matrix oa::FnMatrix::groupedLinearMBiasBwd(const Matrix & inDOut, const Matrix & inOffsets, oa::I32 inNumExperts)

GroupedLinearMBwdResult oa::FnMatrix::groupedLinearMBwd(const Matrix & inDOut, const Matrix & inX, const Matrix & inWeight, const Matrix & inOffsets)

Matrix oa::FnMatrix::gruCellLinear(const Matrix & inGatesI, const Matrix & inHidden, const Matrix & inWeightHh, oa::I32 inHiddenSize, oa::U32 inTimeOffset, oa::U32 inBatchStride, const Matrix & inBiasHh = Matrix{})

Matrix oa::FnMatrix::gruCellPointwise(const Matrix & inGatesI, const Matrix & inGatesH, const Matrix & inHidden, oa::I32 inHiddenSize, oa::U32 inTimeOffset = 0, oa::U32 inBatchStride = 1)

GruCellPointwiseBwdResult oa::FnMatrix::gruCellPointwiseBwd(const Matrix & inGatesI, const Matrix & inGatesH, const Matrix & inHidden, const Matrix & inGradOutput, oa::I32 inHiddenSize, oa::U32 inTimeOffset = 0, oa::U32 inBatchStride = 1)

GruScanResult oa::FnMatrix::gruScan(const Matrix & inGatesI, const Matrix & inWeightHh, oa::I32 inHiddenSize, oa::I32 inSeqLen, oa::I32 inBatch, const Matrix & inBiasHh = Matrix{})

GruScanBwdResult oa::FnMatrix::gruScanBwd(const Matrix & inDOut, const Matrix & inGatesI, const Matrix & inHPrev, const Matrix & inWeightHh, oa::I32 inHiddenSize, oa::I32 inSeqLen, oa::I32 inBatch, const Matrix & inBiasHh = Matrix{})

Matrix oa::FnMatrix::heavyTailActivation(const Matrix & inSelf)

Matrix oa::FnMatrix::im2Col1d(const Matrix & inX, oa::I32 inK, oa::I32 inStride, oa::I32 inPadding, oa::I32 inDilation = 1)

oa::Matrix oa::FnMatrix::layerNorm(const oa::Matrix & inSelf, const oa::Matrix & inWeight, const oa::Matrix & inBias, oa::F32 inEps)

LayerNormBwdResult oa::FnMatrix::layerNormBwd(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, const Matrix & inOut, const Matrix & inMean, const Matrix & inRstd, const Matrix & inGradOutput, oa::F32 inEps = 1e-5F)

oa::Matrix oa::FnMatrix::leakyRelu(const oa::Matrix & inA, oa::F32 inAlpha)

Matrix oa::FnMatrix::leakyReluBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput, oa::F32 inAlpha = 0.01f)

Matrix oa::FnMatrix::linear(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias = Matrix{})

Matrix oa::FnMatrix::linearDataBwd(const Matrix & inGradOutput, const Matrix & inWeight)

Matrix oa::FnMatrix::linearDataReluBwd(const Matrix & inGradOutput, const Matrix & inWeight, const Matrix & inActivation)

Matrix oa::FnMatrix::linearGelu(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias)

Matrix oa::FnMatrix::linearRelu(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias)

Matrix oa::FnMatrix::linearReluBwdData(const Matrix & inGradOutput, const Matrix & inWeight, const Matrix & inActivation)

Matrix oa::FnMatrix::linearSilu(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias)

LinearWeightBiasBwdResult oa::FnMatrix::linearWeightBiasBwd(const Matrix & inInput, const Matrix & inGradOutput)

Matrix oa::FnMatrix::linearWeightBwd(const Matrix & inInput, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::log(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::logSoftmax(const oa::Matrix & inA, oa::I32 inDim = -1)

Matrix oa::FnMatrix::logSoftmaxBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput, oa::I32 inDim = -1)

Matrix oa::FnMatrix::mamba3Mimo(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const Matrix & inMimoX, const Matrix & inMimoZ, const Matrix & inMimoO, const Matrix & inNormWeight, const SsmConfig & inConfig)

Mamba3MimoBwdResult oa::FnMatrix::mamba3MimoBwd(const Matrix & inDOut, const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const Matrix & inMimoX, const Matrix & inMimoZ, const Matrix & inMimoO, const Matrix & inNormWeight, const SsmConfig & inConfig)

Matrix oa::FnMatrix::mamba3MimoStep(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const Matrix & inMimoX, const Matrix & inMimoZ, const Matrix & inMimoO, const Matrix & inNormWeight, const Matrix & inSsmState, const Matrix & inAngleState, const Matrix & inKState, const Matrix & inVState, const SsmConfig & inConfig)

Mamba3PreprocessResult oa::FnMatrix::mamba3Preprocess(const Matrix & inProjected, const Matrix & inDtBias, const Mamba3PreprocessConfig & inConfig)

Mamba3PreprocessBwdResult oa::FnMatrix::mamba3PreprocessBwd(const Matrix & inProjected, const Matrix & inDtBias, const Matrix & inDZ, const Matrix & inDX, const Matrix & inDBh, const Matrix & inDCh, const Matrix & inDDT, const Matrix & inDADT, const Matrix & inDTrap, const Matrix & inDAngle, const Mamba3PreprocessConfig & inConfig)

Matrix oa::FnMatrix::mamba3Siso(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const SsmConfig & inConfig)

SsmBwdResult oa::FnMatrix::mamba3SisoBwd(const Matrix & inDOut, const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const SsmConfig & inConfig)

Matrix oa::FnMatrix::mamba3SisoStep(const Matrix & inC, const Matrix & inB, const Matrix & inX, const Matrix & inZ, const Matrix & inAdt, const Matrix & inDt, const Matrix & inTrap, const Matrix & inAngle, const Matrix & inCBias, const Matrix & inBBias, const Matrix & inD, const Matrix & inSsmState, const Matrix & inAngleState, const Matrix & inKState, const Matrix & inVState, const SsmConfig & inConfig)

Matrix oa::FnMatrix::maskedCategoricalAccuracyCount(const Matrix & inLogits, const Matrix & inLabels, const Matrix & inMask)

oa::Matrix oa::FnMatrix::matMulNt(const oa::Matrix & inA, const oa::Matrix & inB, oa::MatMulPrecision inPrecision = oa::MatMulPrecision::Auto)

oa::Matrix oa::FnMatrix::max(const oa::Matrix & inA, oa::I32 inDim = -1)

Matrix oa::FnMatrix::maxBwd(const Matrix & inInput, const Matrix & inMaxValue, const Matrix & inGradOutput)

MaxPool2dResult oa::FnMatrix::maxPool2d(const Matrix & inX, oa::I32 inKernelSize, oa::I32 inStride, oa::I32 inPadding)

Matrix oa::FnMatrix::maxPool2dBwd(const Matrix & inX, const Matrix & inIndices, const Matrix & inGradOutput, oa::I32 inKernelSize, oa::I32 inStride, oa::I32 inPadding)

oa::Matrix oa::FnMatrix::mean(const oa::Matrix & inA, oa::I32 inDim = -1)

Matrix oa::FnMatrix::mergeHeads(const Matrix & inX, oa::I32 inBatch, oa::I32 inSeqLen, oa::I32 inNumHeads)

oa::Matrix oa::FnMatrix::mish(const oa::Matrix & inA)

Matrix oa::FnMatrix::mishBwd(const Matrix & inInput, const Matrix & inGradOutput)

Matrix oa::FnMatrix::moeCombine(const Matrix & inPacked, const Matrix & inRouteGate, const Matrix & inInverse, const Matrix & inPackedSlot)

MoeCombineBwdResult oa::FnMatrix::moeCombineBwd(const Matrix & inDOut, const Matrix & inPacked, const Matrix & inRouteGate, const Matrix & inInverse, const Matrix & inPackedSlot)

MoeExpertPlan oa::FnMatrix::moeExpertPlan(const Matrix & inExpertIndices, oa::I32 inNumExperts)

Matrix oa::FnMatrix::moeGather(const Matrix & inSelf, const Matrix & inIndices, const Matrix & inInverse)

Matrix oa::FnMatrix::moeGatherBwd(const Matrix & inSource, const Matrix & inInverse, oa::I32 inOutRows)

Matrix oa::FnMatrix::moeRouteWeights(const Matrix & inProbs, const Matrix & inExpertIndices)

Matrix oa::FnMatrix::moeRouteWeightsBwd(const Matrix & inDOut, const Matrix & inProbs, const Matrix & inExpertIndices, const Matrix & inRouteWeights)

void oa::FnMatrix::moeRoutingBiasUpdate(const Matrix & inSelectionMask, Matrix & inOutBias, oa::I32 inExpertsPerToken, oa::F32 inGamma)

oa::Matrix oa::FnMatrix::mul(const oa::Matrix & inA, const oa::Matrix & inB)

void oa::FnMatrix::mulInPlace(Matrix & inSelf, const Matrix & inOther)

oa::Matrix oa::FnMatrix::neg(const oa::Matrix & inA)

Matrix oa::FnMatrix::ones(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

oa::Matrix oa::FnMatrix::philoxNormal(const oa::Matrix & inA, oa::F32 inMean = 0.0F, oa::F32 inStddev = 1.0F, oa::U64 inSeed = 0)

oa::Matrix oa::FnMatrix::philoxUniform(const oa::Matrix & inA, oa::F32 inLow = 0.0F, oa::F32 inHigh = 1.0F, oa::U64 inSeed = 0)

oa::Matrix oa::FnMatrix::pow(const oa::Matrix & inA, oa::F32 inExponent)

oa::QuantMatrix oa::FnMatrix::quantize(const oa::Matrix & inInput, oa::Quantization inQuantization)

Matrix oa::FnMatrix::rand(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::randGlorotUniform(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::randKaimingUniform(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::randN(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Matrix oa::FnMatrix::randXavier(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

oa::Matrix oa::FnMatrix::reciprocal(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::relu(const oa::Matrix & inA)

Matrix oa::FnMatrix::reluBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput)

Matrix oa::FnMatrix::repeatInterleave(const Matrix & inA, oa::I32 inRepeats, oa::I32 inDim)

Matrix oa::FnMatrix::repeatInterleaveBwd(const Matrix & inGradOut, MatrixShape inInputShape, oa::I32 inRepeats, oa::I32 inDim)

oa::Matrix oa::FnMatrix::reshape(const oa::Matrix & inA, oa::MatrixShape inShape)

ResidualRmsNormResult oa::FnMatrix::residualRmsNorm(const Matrix & inA, const Matrix & inB, const Matrix & inWeight, oa::F32 inEps)

oa::Matrix oa::FnMatrix::rmsNorm(const oa::Matrix & inSelf, const oa::Matrix & inWeight, oa::F32 inEps)

RmsNormBwdResult oa::FnMatrix::rmsNormBwd(const Matrix & inX, const Matrix & inWeight, const Matrix & inGradOutput, oa::F32 inEps = 1e-5F)

Matrix oa::FnMatrix::rmsNormGated(const Matrix & inSelf, const Matrix & inWeight, const Matrix & inBias, const Matrix & inZ, oa::F32 inEps, bool inNormBeforeGate = true)

RmsNormGatedBwdResult oa::FnMatrix::rmsNormGatedBwd(const Matrix & inX, const Matrix & inWeight, const Matrix & inBias, const Matrix & inZ, const Matrix & inGradOutput, oa::F32 inEps)

Matrix oa::FnMatrix::rnnCellLinear(const Matrix & inGatesI, const Matrix & inHidden, const Matrix & inWeightHh, oa::U32 inTimeOffset = 0, oa::U32 inBatchStride = 1, const Matrix & inBiasHh = Matrix{})

Matrix oa::FnMatrix::rnnCellPointwise(const Matrix & inGatesI, const Matrix & inGatesH)

RnnCellPointwiseBwdResult oa::FnMatrix::rnnCellPointwiseBwd(const Matrix & inGatesI, const Matrix & inGatesH, const Matrix & inGradOutput, oa::I32 inHiddenSize, oa::U32 inTimeOffset = 0, oa::U32 inBatchStride = 1)

RnnScanResult oa::FnMatrix::rnnScan(const Matrix & inGatesI, const Matrix & inWeightHh, oa::I32 inHiddenSize, oa::I32 inSeqLen, oa::I32 inBatch, const Matrix & inBiasHh = Matrix{})

RnnScanBwdResult oa::FnMatrix::rnnScanBwd(const Matrix & inDOut, const Matrix & inGatesI, const Matrix & inHPrev, const Matrix & inWeightHh, oa::I32 inHiddenSize, oa::I32 inSeqLen, oa::I32 inBatch, const Matrix & inBiasHh = Matrix{})

Matrix oa::FnMatrix::sampleLogits(const Matrix & inLogits, oa::F32 inTemperature = 1.0F, oa::I32 inTopK = 0, oa::F32 inTopP = 1.0F, oa::U64 inSeed = 0)

oa::F32 oa::FnMatrix::scalar(const Matrix & inSrc)

oa::Matrix oa::FnMatrix::scale(const oa::Matrix & inA, oa::F32 inScalar)

void oa::FnMatrix::scaleInPlace(Matrix & inSelf, oa::F32 inScalar)

Matrix oa::FnMatrix::scatterAddRows(const Matrix & inSource, const Matrix & inIndices, oa::I32 inOutRows)

Matrix oa::FnMatrix::scatterRows(const Matrix & inSelf, const Matrix & inSource, const Matrix & inRowMap, const Matrix & inCount)

Matrix oa::FnMatrix::scatterRowsBwdSource(const Matrix & inGradOut, const Matrix & inRowMap, const Matrix & inCount)

void oa::FnMatrix::setRngSeed(oa::U64 inSeed)

oa::Matrix oa::FnMatrix::sigmoid(const oa::Matrix & inA)

Matrix oa::FnMatrix::sigmoidBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::silu(const oa::Matrix & inA)

Matrix oa::FnMatrix::siluBwd(const Matrix & inInput, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::siluMul(const oa::Matrix & inA, oa::U32 inIntermediateSize)

Matrix oa::FnMatrix::siluMulBwd(const Matrix & inInput, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::sin(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::slice(const oa::Matrix & inSelf, oa::I32 inDim, oa::I64 inStart, oa::I64 inEnd)

Matrix oa::FnMatrix::sliceBwd(MatrixShape inInputShape, oa::I32 inDim, oa::I64 inStart, oa::I64 inEnd, const Matrix & inDOut)

oa::Matrix oa::FnMatrix::softmax(const oa::Matrix & inA, oa::I32 inDim = -1)

Matrix oa::FnMatrix::softmaxBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput, oa::I32 inDim = -1)

Matrix oa::FnMatrix::softmaxScaledMasked(const Matrix & inScores, const Matrix & inMask, oa::F32 inScale)

Matrix oa::FnMatrix::softmaxScaledMaskedBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput, oa::F32 inScale)

oa::Matrix oa::FnMatrix::softplus(const oa::Matrix & inA)

Matrix oa::FnMatrix::softplusBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput)

oa::Vector<Matrix> oa::FnMatrix::split(const Matrix & inSelf, oa::Span<oa::I64> inSizes, oa::I32 inDim = 0)

Matrix oa::FnMatrix::splitHeads(const Matrix & inX, oa::I32 inBatch, oa::I32 inSeqLen, oa::I32 inNumHeads)

oa::Matrix oa::FnMatrix::sqrt(const oa::Matrix & inA)

oa::Matrix oa::FnMatrix::sub(const oa::Matrix & inA, const oa::Matrix & inB)

void oa::FnMatrix::subInPlace(Matrix & inSelf, const Matrix & inOther)

oa::Matrix oa::FnMatrix::subScalar(const oa::Matrix & inA, oa::F32 inScalar)

void oa::FnMatrix::subScalarInPlace(Matrix & inSelf, oa::F32 inScalar)

oa::Matrix oa::FnMatrix::sum(const oa::Matrix & inA, oa::I32 inDim = -1)

oa::Matrix oa::FnMatrix::swiglu(const oa::Matrix & inA, const oa::Matrix & inB)

SwigluBwdResult oa::FnMatrix::swigluBwd(const Matrix & inGate, const Matrix & inUp, const Matrix & inOut, const Matrix & inGradOutput)

oa::Matrix oa::FnMatrix::tanh(const oa::Matrix & inA)

Matrix oa::FnMatrix::tanhBwd(const Matrix & inForwardOutput, const Matrix & inGradOutput)

TopKResult oa::FnMatrix::topK(const Matrix & inA, oa::I32 inK, oa::I32 inDim = -1)

Matrix oa::FnMatrix::topKMask(const Matrix & inIndices, oa::I32 inNumExperts)

oa::Matrix oa::FnMatrix::transpose(const oa::Matrix & inA, oa::I32 inDim0 = -2, oa::I32 inDim1 = -1)

Matrix oa::FnMatrix::upsampleBwd(const Matrix & inInput, const Matrix & inDOut, oa::I32 inScaleFactor, bool inIsBilinear)

VqAssignResult oa::FnMatrix::vqAssign(const Matrix & inZe, const Matrix & inCodebook)

void oa::FnMatrix::vqEmaUpdate(const Matrix & inZe, const Matrix & inIdx, Matrix & ioEmbedSum, Matrix & ioClusterSize, Matrix & outCodebook, oa::F32 inDecay, oa::F32 inEps, oa::F32 inDeadThreshold, oa::U32 inSeed, bool inNormalize)

oa::ScalarType oa::FnMatrix::weightDtype()

Matrix oa::FnMatrix::zeros(MatrixShape inShape, oa::ScalarType inDtype = weightDtype())

Function Documentation

oa::Matrix oa::FnMatrix::abs(
	const oa::Matrix & inA
)

abs: Element-wise absolute value: out = |A|.

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::add(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

add: Element-wise addition: out = A + B.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::addInPlace(
	Matrix & inSelf,
	const Matrix & inOther
)

in-place operations

Parameters

inSelf

Matrix &

—

inOther

const Matrix &

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::addScalar(
	const oa::Matrix & inA,
	oa::F32 inScalar
)

addScalar: Scalar addition: out = A + Scalar.

Parameters

inA

const oa::Matrix &

—

inScalar

oa::F32

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::addScalarInPlace(
	Matrix & inSelf,
	oa::F32 inScalar
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inScalar

oa::F32

—

Returns

void

The declared return value.

oa::I64 oa::FnMatrix::argmax(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

argmax: Find index of maximum value: out = argmax(A).

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::I64

The declared return value.

Matrix oa::FnMatrix::avgPool2d(
	const Matrix & inX,
	oa::I32 inKernelSize,
	oa::I32 inStride,
	oa::I32 inPadding
)

--- Pooling ---

Parameters

inX

const Matrix &

—

inKernelSize

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::avgPool2dBwd(
	const Matrix & inX,
	const Matrix & inGradOutput,
	oa::I32 inKernelSize,
	oa::I32 inStride,
	oa::I32 inPadding
)

avgPool2dBwd: backward for 2D average pooling.

Parameters

inX

const Matrix &

—

inGradOutput

const Matrix &

—

inKernelSize

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

Returns

Matrix

The declared return value.

BatchNorm2dBwdResult oa::FnMatrix::batchNorm2dBwd(
	const Matrix & inX,
	const Matrix & inGamma,
	const Matrix & inBeta,
	const Matrix & inMean,
	const Matrix & inVar,
	const Matrix & inOut,
	const Matrix & inDOut,
	oa::F32 inEps,
	bool inIsTraining
)

batchNorm2dBwd: backward pass for BatchNorm2d.

Parameters

inX

const Matrix &

—

inGamma

const Matrix &

—

inBeta

const Matrix &

—

inMean

const Matrix &

—

inVar

const Matrix &

—

inOut

const Matrix &

—

inDOut

const Matrix &

—

inEps

oa::F32

—

inIsTraining

bool

—

Returns

BatchNorm2dBwdResult

The declared return value.

oa::Matrix oa::FnMatrix::biasAdd(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

biasAdd: Add bias vector: out = A + Bias (broadcasted).

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::bmm(
	const Matrix & inA,
	const Matrix & inB
)

bmm: per-batch matrix multiply, A[N,M,K] @ B[N,K,P] = out[N,M,P].

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::bmmNt(
	const Matrix & inA,
	const Matrix & inB
)

bmmNt: per-batch matrix multiply with transposed right operand storage, A[N,M,K] @ B[N,P,K]^T = out[N,M,P].

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::cast(
	const Matrix & inSrc,
	oa::ScalarType inDtype
)

--- dtype cast --- cast: allocate a new matrix of inDtype and convert inSrc into it. castInto: convert inSrc into the pre-allocated outDst in place.

Parameters

inSrc

const Matrix &

—

inDtype

oa::ScalarType

—

Returns

Matrix

The declared return value.

void oa::FnMatrix::castInto(
	const Matrix & inSrc,
	Matrix & outDst
)

No public source comment is attached to this declaration.

Parameters

inSrc

const Matrix &

—

outDst

Matrix &

—

Returns

void

The declared return value.

Matrix oa::FnMatrix::categoricalAccuracyCount(
	const Matrix & inLogits,
	const Matrix & inLabels
)

No public source comment is attached to this declaration.

Parameters

inLogits

const Matrix &

—

inLabels

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::causalMask(
	oa::I64 inSeqLen
)
Matrix oa::FnMatrix::causalMask(
	const Matrix & inScores
)

No public source comment is attached to this declaration.

Parameters

inSeqLen

oa::I64

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::causalMaskBwd(
	const Matrix & inGradOut
)

No public source comment is attached to this declaration.

Parameters

inGradOut

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::channelNorm(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::I32 inBatch,
	oa::I32 inChannels,
	oa::I32 inSeqLen,
	oa::F32 inEps
)

channelNorm: fused LayerNorm over the channel axis of [B,C,T] without transposing. Replaces Transpose+LayerNorm+transpose (3 dispatches) with 1.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inBatch

oa::I32

—

inChannels

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

—

Returns

Matrix

The declared return value.

ChannelNormBwdResult oa::FnMatrix::channelNormBwd(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inGradOutput,
	oa::I32 inBatch,
	oa::I32 inChannels,
	oa::I32 inSeqLen,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inGradOutput

const Matrix &

—

inBatch

oa::I32

—

inChannels

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

—

Returns

ChannelNormBwdResult

The declared return value.

Matrix oa::FnMatrix::channelNormRelu(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::I32 inBatch,
	oa::I32 inChannels,
	oa::I32 inSeqLen,
	oa::F32 inEps
)

channelNormRelu: fused channelNorm + ReLU on [B,C,T].

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inBatch

oa::I32

—

inChannels

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

—

Returns

Matrix

The declared return value.

ChannelNormBwdResult oa::FnMatrix::channelNormReluBwd(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inFwdOut,
	const Matrix & inGradOutput,
	oa::I32 inBatch,
	oa::I32 inChannels,
	oa::I32 inSeqLen,
	oa::F32 inEps
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inFwdOut

const Matrix &

—

inGradOutput

const Matrix &

—

inBatch

oa::I32

—

inChannels

oa::I32

—

inSeqLen

oa::I32

—

inEps

oa::F32

—

Returns

ChannelNormBwdResult

The declared return value.

oa::Matrix oa::FnMatrix::clampMax(
	const oa::Matrix & inA,
	oa::F32 inMax
)

clampMax: Element-wise clamp max: out = min(A, Max).

Parameters

inA

const oa::Matrix &

—

inMax

oa::F32

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::clampMin(
	const oa::Matrix & inA,
	oa::F32 inMin
)

clampMin: Element-wise clamp min: out = max(A, Min).

Parameters

inA

const oa::Matrix &

—

inMin

oa::F32

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::col2Im1d(
	const Matrix & inDCols,
	oa::I32 inN,
	oa::I32 inC,
	oa::I32 inL,
	oa::I32 inK,
	oa::I32 inStride,
	oa::I32 inPadding,
	oa::I32 inDilation,
	oa::I32 inOutL
)

col2Im1d: backward of im2Col1d — fold a column-matrix gradient [N*outL, inC*K] back into input shape [N, inC, L], accumulating over overlapping windows.

Parameters

inDCols

const Matrix &

—

inN

oa::I32

—

inC

oa::I32

—

inL

oa::I32

—

inK

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

inDilation

oa::I32

—

inOutL

oa::I32

—

Returns

Matrix

The declared return value.

CompactRowsResult oa::FnMatrix::compactRows(
	const Matrix & inSelf,
	const Matrix & inMask
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

inMask

const Matrix &

—

Returns

CompactRowsResult

The declared return value.

Matrix oa::FnMatrix::compactRowsBwd(
	const Matrix & inGradOut,
	const Matrix & inRowMap,
	const Matrix & inCount,
	MatrixShape inInputShape
)
Matrix oa::FnMatrix::compactRowsBwd(
	const Matrix & inGradOut,
	const Matrix & inRowMap,
	const Matrix & inCount,
	const Matrix & inDispatchArgs,
	MatrixShape inInputShape
)

No public source comment is attached to this declaration.

Parameters

inGradOut

const Matrix &

—

inRowMap

const Matrix &

—

inCount

const Matrix &

—

inInputShape

MatrixShape

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::concat(
	oa::Span<Matrix> inInputs,
	oa::I32 inDim = 0
)

No public source comment is attached to this declaration.

Parameters

inInputs

oa::Span<Matrix>

—

inDim

oa::I32

Default: 0

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::conv1dBwdData(
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding,
	oa::U32 inDilation,
	const MatrixShape & inInputShape
)

conv1dBwdData: backward for 1D convolution (input gradient).

Parameters

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inDilation

oa::U32

—

inInputShape

const MatrixShape &

—

Returns

Matrix

The declared return value.

Conv1dBwdWeightResult oa::FnMatrix::conv1dBwdWeight(
	const Matrix & inInput,
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding,
	oa::U32 inDilation
)

conv1dBwdWeight: fused weight and bias gradient for 1D convolution.

Parameters

inInput

const Matrix &

—

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inDilation

oa::U32

—

Returns

Conv1dBwdWeightResult

The declared return value.

Matrix oa::FnMatrix::conv1dGemm(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::I32 inStride = 1,
	oa::I32 inPadding = 0,
	oa::I32 inDilation = 1
)

conv1dGemm: 1-D convolution executed as im2col + a single matmul. inX [N, inC, L], inWeight [outC, inC, K], inBias [outC] -> [N, outC, outL]

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inStride

oa::I32

Default: 1

inPadding

oa::I32

Default: 0

inDilation

oa::I32

Default: 1

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::conv1dReluGemm(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::I32 inStride = 1,
	oa::I32 inPadding = 0,
	oa::I32 inDilation = 1
)

conv1dReluGemm: conv1dGemm with the ReLU folded into the GEMM bias epilogue.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inStride

oa::I32

Default: 1

inPadding

oa::I32

Default: 0

inDilation

oa::I32

Default: 1

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::conv2d(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::U32 inStride,
	oa::U32 inPadding,
	oa::U32 inGroups = 1
)

--- Conv2d ---

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inGroups

oa::U32

Default: 1

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::conv2dBwdData(
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding,
	const MatrixShape & inInputShape,
	oa::U32 inGroups = 1
)

conv2dBwdData: backward for 2D convolution (input gradient).

Parameters

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inInputShape

const MatrixShape &

—

inGroups

oa::U32

Default: 1

Returns

Matrix

The declared return value.

Conv2dBwdWeightResult oa::FnMatrix::conv2dBwdWeight(
	const Matrix & inInput,
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding,
	oa::U32 inGroups = 1
)

conv2dBwdWeight: fused weight and bias gradient for 2D convolution.

Parameters

inInput

const Matrix &

—

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inGroups

oa::U32

Default: 1

Returns

Conv2dBwdWeightResult

The declared return value.

Matrix oa::FnMatrix::convTranspose2d(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	oa::U32 inStride,
	oa::U32 inPadding
)

convTranspose2d: 2D transposed convolution (learnable upsampling). input: [N, inC, H, W], weight: [inC, outC, K, K], Bias: [outC] output: [N, outC, H_out, W_out] where H_out = (H - 1) * S - 2P + K.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::convTranspose2dBwdData(
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding,
	const MatrixShape & inInputShape
)

convTranspose2dBwdData: backward for 2D transposed convolution (input gradient).

Parameters

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

inInputShape

const MatrixShape &

—

Returns

Matrix

The declared return value.

ConvTranspose2dBwdWeightResult oa::FnMatrix::convTranspose2dBwdWeight(
	const Matrix & inInput,
	const Matrix & inDOut,
	const Matrix & inWeight,
	oa::U32 inStride,
	oa::U32 inPadding
)

convTranspose2dBwdWeight: fused weight and bias gradient for 2D transposed convolution.

Parameters

inInput

const Matrix &

—

inDOut

const Matrix &

—

inWeight

const Matrix &

—

inStride

oa::U32

—

inPadding

oa::U32

—

Returns

ConvTranspose2dBwdWeightResult

The declared return value.

oa::Matrix oa::FnMatrix::copy(
	const oa::Matrix & inA
)

copy: Copy tensor data: out = A.

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Status oa::FnMatrix::copyToHost(
	const Matrix & inSrc,
	void * outHost,
	oa::U64 inBytes
)

--- Transfer --- Copy device matrix data to host memory. inBytes must be >= inSrc.byteSize().

Parameters

inSrc

const Matrix &

—

outHost

void *

—

inBytes

oa::U64

—

Returns

oa::Status

The declared return value.

oa::Matrix oa::FnMatrix::cos(
	const oa::Matrix & inA
)

cos: Element-wise cosine: out = cos(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::dequantize(
	const oa::QuantMatrix & inInput
)

Materialize a Float32 matrix with the logical shape retained by inInput.

Parameters

inInput

const oa::QuantMatrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::detach(
	const Matrix & inSelf
)

detach: stop-gradient. Returns a view that SHARES inSelf's device buffer but carries no autograd linkage (leaf, requiresGrad=false), so backward terminates here. Metadata-only: no kernel, no copy. This is the primitive the straight- through estimator needs.

Parameters

inSelf

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::div(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

div: Element-wise division: out = A / B.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::divInPlace(
	Matrix & inSelf,
	const Matrix & inOther
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inOther

const Matrix &

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::divScalar(
	const oa::Matrix & inA,
	oa::F32 inScalar
)

divScalar: Scalar division: out = A / Scalar.

Parameters

inA

const oa::Matrix &

—

inScalar

oa::F32

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::divScalarInPlace(
	Matrix & inSelf,
	oa::F32 inScalar
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inScalar

oa::F32

—

Returns

void

The declared return value.

Matrix oa::FnMatrix::dropout(
	const Matrix & inA,
	oa::F32 inP,
	oa::U64 inSeed = 0
)

Inverted dropout. inP must be [0,1).

Parameters

inA

const Matrix &

—

inP

oa::F32

—

inSeed

oa::U64

Default: 0

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::elu(
	const oa::Matrix & inA,
	oa::F32 inAlpha
)

elu: ELU activation: out = A if A > 0 else Alpha * (exp(A) - 1).

Parameters

inA

const oa::Matrix &

—

inAlpha

oa::F32

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::eluBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput,
	oa::F32 inAlpha = 1.0f
)

eluBwd: backward pass for ELU activation. Computes: dInput = dOutput * (x > 0 ? 1 : alpha * exp(x))

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

inAlpha

oa::F32

Default: 1.0f

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::empty(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype(),
	oa::MemoryPlacement inPlacement = oa::MemoryPlacement::Auto
)

--- Factory functions ---

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

inPlacement

oa::MemoryPlacement

Default: oa::MemoryPlacement::Auto

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::empyrealmAdt(
	const Matrix & inDdA,
	const Matrix & inDt,
	oa::F32 inAFloor
)

--- empyrealm operations ---

Parameters

inDdA

const Matrix &

—

inDt

const Matrix &

—

inAFloor

oa::F32

—

Returns

Matrix

The declared return value.

EmpyrealmAdtBwdResult oa::FnMatrix::empyrealmAdtBwd(
	const Matrix & inDOut,
	const Matrix & inDdA,
	const Matrix & inDt,
	oa::F32 inAFloor
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inDdA

const Matrix &

—

inDt

const Matrix &

—

inAFloor

oa::F32

—

Returns

EmpyrealmAdtBwdResult

The declared return value.

Matrix oa::FnMatrix::empyrealmDt(
	const Matrix & inX,
	oa::F32 inDtMin,
	oa::F32 inDtMax
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inDtMin

oa::F32

—

inDtMax

oa::F32

—

Returns

Matrix

The declared return value.

EmpyrealmDtAdtResult oa::FnMatrix::empyrealmDtAdt(
	const Matrix & inDtRaw,
	const Matrix & inDdA,
	oa::F32 inDtMin,
	oa::F32 inDtMax,
	oa::F32 inAFloor
)

No public source comment is attached to this declaration.

Parameters

inDtRaw

const Matrix &

—

inDdA

const Matrix &

—

inDtMin

oa::F32

—

inDtMax

oa::F32

—

inAFloor

oa::F32

—

Returns

EmpyrealmDtAdtResult

The declared return value.

Matrix oa::FnMatrix::empyrealmDtBwd(
	const Matrix & inDOut,
	const Matrix & inX,
	oa::F32 inDtMin,
	oa::F32 inDtMax
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inX

const Matrix &

—

inDtMin

oa::F32

—

inDtMax

oa::F32

—

Returns

Matrix

The declared return value.

Mamba3PreprocessResult oa::FnMatrix::empyrealmPreprocess(
	const Matrix & inProjected,
	const Matrix & inDtBias,
	const Mamba3PreprocessConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inProjected

const Matrix &

—

inDtBias

const Matrix &

—

inConfig

const Mamba3PreprocessConfig &

—

Returns

Mamba3PreprocessResult

The declared return value.

Mamba3PreprocessBwdResult oa::FnMatrix::empyrealmPreprocessBwd(
	const Matrix & inProjected,
	const Matrix & inDtBias,
	const Matrix & inDZ,
	const Matrix & inDX,
	const Matrix & inDBh,
	const Matrix & inDCh,
	const Matrix & inDDT,
	const Matrix & inDADT,
	const Matrix & inDTrap,
	const Matrix & inDAngle,
	const Mamba3PreprocessConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inProjected

const Matrix &

—

inDtBias

const Matrix &

—

inDZ

const Matrix &

—

inDX

const Matrix &

—

inDBh

const Matrix &

—

inDCh

const Matrix &

—

inDDT

const Matrix &

—

inDADT

const Matrix &

—

inDTrap

const Matrix &

—

inDAngle

const Matrix &

—

inConfig

const Mamba3PreprocessConfig &

—

Returns

Mamba3PreprocessBwdResult

The declared return value.

Matrix oa::FnMatrix::empyrealmSiso(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

SsmBwdResult oa::FnMatrix::empyrealmSisoBwd(
	const Matrix & inDOut,
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

SsmBwdResult

The declared return value.

Matrix oa::FnMatrix::empyrealmSisoStep(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const Matrix & inSsmState,
	const Matrix & inAngleState,
	const Matrix & inKState,
	const Matrix & inVState,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inSsmState

const Matrix &

—

inAngleState

const Matrix &

—

inKState

const Matrix &

—

inVState

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::equal(
	const Matrix & inA,
	oa::F32 inValue
)

No public source comment is attached to this declaration.

Parameters

inA

const Matrix &

—

inValue

oa::F32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::exp(
	const oa::Matrix & inA
)

exp: Element-wise exponential: out = exp(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::fill(
	const oa::MatrixShape & inShape,
	oa::F32 inValue
)

fill: Create a Float32 matrix filled with a constant value.

Parameters

inShape

const oa::MatrixShape &

—

inValue

oa::F32

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::fillInPlace(
	oa::Matrix & inSelf,
	oa::F32 inValue
)

fillInPlace: Replace every value in an existing matrix.

Parameters

inSelf

oa::Matrix &

—

inValue

oa::F32

—

Returns

void

The declared return value.

Matrix oa::FnMatrix::flashAttentionCausal(
	const Matrix & inQ,
	const Matrix & inK,
	const Matrix & inV,
	oa::F32 inScale
)

flashAttentionCausal: IO-aware causal scaled dot-product attention. Q/K/V and output use contiguous [batchHeads, sequence, headDim] storage.

Parameters

inQ

const Matrix &

—

inK

const Matrix &

—

inV

const Matrix &

—

inScale

oa::F32

—

Returns

Matrix

The declared return value.

FlashAttentionBwdResult oa::FnMatrix::flashAttentionCausalBwd(
	const Matrix & inQ,
	const Matrix & inK,
	const Matrix & inV,
	const Matrix & inOutput,
	const Matrix & inLogSumExp,
	const Matrix & inGradOutput,
	oa::F32 inScale
)

flashAttentionCausalBwd: explicit adjoint for the FlashAttention autograd node.

Parameters

inQ

const Matrix &

—

inK

const Matrix &

—

inV

const Matrix &

—

inOutput

const Matrix &

—

inLogSumExp

const Matrix &

—

inGradOutput

const Matrix &

—

inScale

oa::F32

—

Returns

FlashAttentionBwdResult

The declared return value.

Matrix oa::FnMatrix::fromBytes(
	oa::Span<const oa::U8> inData,
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inData

oa::Span<const oa::U8>

—

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::fromInt32(
	oa::Span<const oa::I32> inData,
	MatrixShape inShape,
	oa::ScalarType inDtype = oa::ScalarType::Int32
)

No public source comment is attached to this declaration.

Parameters

inData

oa::Span<const oa::I32>

—

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: oa::ScalarType::Int32

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::full(
	MatrixShape inShape,
	oa::F64 inValue,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inValue

oa::F64

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::gather(
	const Matrix & inSelf,
	const Matrix & inIndices
)

Indexing

Parameters

inSelf

const Matrix &

—

inIndices

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::gatherBwd(
	const Matrix & inIndices,
	const Matrix & inGradOutput,
	oa::I32 inVocabSize,
	oa::I32 inEmbedDim
)

No public source comment is attached to this declaration.

Parameters

inIndices

const Matrix &

—

inGradOutput

const Matrix &

—

inVocabSize

oa::I32

—

inEmbedDim

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::gatherLastDim(
	const Matrix & inSelf,
	const Matrix & inIndices
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

inIndices

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::gatherLastDimBwd(
	const Matrix & inGradOut,
	const Matrix & inIndices,
	oa::I32 inInputWidth
)

No public source comment is attached to this declaration.

Parameters

inGradOut

const Matrix &

—

inIndices

const Matrix &

—

inInputWidth

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::geglu(
	const oa::Matrix & inA,
	oa::U32 inIntermediateSize
)

geglu: GEGLU activation: out = A[:N] * GELU(A[N:]).

Parameters

inA

const oa::Matrix &

—

inIntermediateSize

oa::U32

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::gegluBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

gegluBwd: backward pass for GEGLU activation.

Parameters

inInput

const Matrix &

forward INPUT (up||gate); up*GELU(gate) is not invertible

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::gelu(
	const oa::Matrix & inA
)

gelu: GELU activation: out = GELU(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::geluBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

geluBwd: backward pass for GELU activation. Computes: dInput = dOutput * gelu'(x)

Parameters

inInput

const Matrix &

forward INPUT x (gelu'(x) is a function of input, not output)

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::greaterEqual(
	const Matrix & inA,
	oa::F32 inValue
)

No public source comment is attached to this declaration.

Parameters

inA

const Matrix &

—

inValue

oa::F32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::groupedGemmM(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inOffsets
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inOffsets

const Matrix &

—

Returns

Matrix

The declared return value.

GroupedGemmMBwdResult oa::FnMatrix::groupedGemmMBwd(
	const Matrix & inDOut,
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inOffsets
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inX

const Matrix &

—

inWeight

const Matrix &

—

inOffsets

const Matrix &

—

Returns

GroupedGemmMBwdResult

The declared return value.

Matrix oa::FnMatrix::groupedLinearM(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	const Matrix & inOffsets
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inOffsets

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::groupedLinearMBiasBwd(
	const Matrix & inDOut,
	const Matrix & inOffsets,
	oa::I32 inNumExperts
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inOffsets

const Matrix &

—

inNumExperts

oa::I32

—

Returns

Matrix

The declared return value.

GroupedLinearMBwdResult oa::FnMatrix::groupedLinearMBwd(
	const Matrix & inDOut,
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inOffsets
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inX

const Matrix &

—

inWeight

const Matrix &

—

inOffsets

const Matrix &

—

Returns

GroupedLinearMBwdResult

The declared return value.

Matrix oa::FnMatrix::gruCellLinear(
	const Matrix & inGatesI,
	const Matrix & inHidden,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::U32 inTimeOffset,
	oa::U32 inBatchStride,
	const Matrix & inBiasHh = Matrix{}
)

gruCellLinear: fused GRU recurrent step — Linear(h, W_hh) + gruCellPointwise. Replaces the per-timestep pair of dispatches with one kernel. The hidden projection required by reverse mode is retained internally rather than exposed as an output parameter.

Parameters

inGatesI

const Matrix &

—

inHidden

const Matrix &

—

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inTimeOffset

oa::U32

—

inBatchStride

oa::U32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::gruCellPointwise(
	const Matrix & inGatesI,
	const Matrix & inGatesH,
	const Matrix & inHidden,
	oa::I32 inHiddenSize,
	oa::U32 inTimeOffset = 0,
	oa::U32 inBatchStride = 1
)

gruCellPointwise: fused GRU pointwise forward. r = sigmoid(gatesI[r] + gatesH[r]) z = sigmoid(gatesI[z] + gatesH[z]) n = tanh(gatesI[n] + r * gatesH[n]) h_new = (1 - z) * n + z * hPrev

Parameters

inGatesI

const Matrix &

[B, 3H] input projection (reset|update|candidate along dim 1) or [B*T, 3H] with inTimeOffset = t*B to index row t without Slice

inGatesH

const Matrix &

[B, 3H] hidden projection

inHidden

const Matrix &

[B, H] previous hidden state

inHiddenSize

oa::I32

inTimeOffset

oa::U32

row offset into inGatesI (in units of 3H), default 0

inBatchStride

oa::U32

Default: 1

Returns

Matrix

The declared return value.

GruCellPointwiseBwdResult oa::FnMatrix::gruCellPointwiseBwd(
	const Matrix & inGatesI,
	const Matrix & inGatesH,
	const Matrix & inHidden,
	const Matrix & inGradOutput,
	oa::I32 inHiddenSize,
	oa::U32 inTimeOffset = 0,
	oa::U32 inBatchStride = 1
)

gruCellPointwiseBwd: fused GRU pointwise backward. Returns gradients w.r.t. gatesI, gatesH and the previous hidden state.

Parameters

inGatesI

const Matrix &

—

inGatesH

const Matrix &

—

inHidden

const Matrix &

—

inGradOutput

const Matrix &

—

inHiddenSize

oa::I32

—

inTimeOffset

oa::U32

row offset into inGatesI / dGatesI (in rows of 3H), default 0

inBatchStride

oa::U32

row stride between batches in inGatesI (T for batch-major, 1 for contiguous), default 1

Returns

GruCellPointwiseBwdResult

The declared return value.

GruScanResult oa::FnMatrix::gruScan(
	const Matrix & inGatesI,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::I32 inSeqLen,
	oa::I32 inBatch,
	const Matrix & inBiasHh = Matrix{}
)

gruScan: whole-sequence GRU recurrent scan in ONE dispatch (one workgroup per batch, looping all timesteps). Mathematically identical to running gruCellLinear for each timestep, but collapses S dispatches into 1. The recurrent weight/bias gradient is computed separately via linearWeightBiasBwd on the saved hPrev.

Parameters

inGatesI

const Matrix &

[B*S, 3H] precomputed input projection (row b*S+t = timestep t)

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inSeqLen

oa::I32

inBatch

oa::I32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

GruScanResult

The declared return value.

GruScanBwdResult oa::FnMatrix::gruScanBwd(
	const Matrix & inDOut,
	const Matrix & inGatesI,
	const Matrix & inHPrev,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::I32 inSeqLen,
	oa::I32 inBatch,
	const Matrix & inBiasHh = Matrix{}
)

gruScanBwd: BPTT recurrence scan (backward of gruScan) in ONE dispatch. Produces gradients w.r.t. the input projection gatesI and the hidden projection gatesH (the latter drives the separate linearWeightBiasBwd weight-grad call).

Parameters

inDOut

const Matrix &

—

inGatesI

const Matrix &

—

inHPrev

const Matrix &

—

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inSeqLen

oa::I32

—

inBatch

oa::I32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

GruScanBwdResult

The declared return value.

Matrix oa::FnMatrix::heavyTailActivation(
	const Matrix & inSelf
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::im2Col1d(
	const Matrix & inX,
	oa::I32 inK,
	oa::I32 inStride,
	oa::I32 inPadding,
	oa::I32 inDilation = 1
)

im2Col1d: unfold a 1-D conv input [N, inC, L] into the GEMM-ready column matrix [N*outL, inC*K]. The building block of conv1dGemm.

Parameters

inX

const Matrix &

—

inK

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

inDilation

oa::I32

Default: 1

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::layerNorm(
	const oa::Matrix & inSelf,
	const oa::Matrix & inWeight,
	const oa::Matrix & inBias,
	oa::F32 inEps
)

layerNorm: Affine layer normalization over the last dimension: out = ((A - mean(A)) / sqrt(var(A) + Eps)) * Weight + Bias.

Parameters

inSelf

const oa::Matrix &

—

inWeight

const oa::Matrix &

—

inBias

const oa::Matrix &

—

inEps

oa::F32

—

Returns

oa::Matrix

The declared return value.

LayerNormBwdResult oa::FnMatrix::layerNormBwd(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	const Matrix & inOut,
	const Matrix & inMean,
	const Matrix & inRstd,
	const Matrix & inGradOutput,
	oa::F32 inEps = 1e-5F
)

layerNormBwd: backward pass for LayerNorm.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inOut

const Matrix &

—

inMean

const Matrix &

—

inRstd

const Matrix &

—

inGradOutput

const Matrix &

—

inEps

oa::F32

Default: 1e-5F

Returns

LayerNormBwdResult

The declared return value.

oa::Matrix oa::FnMatrix::leakyRelu(
	const oa::Matrix & inA,
	oa::F32 inAlpha
)

leakyRelu: Leaky ReLU activation: out = max(Alpha * A, A).

Parameters

inA

const oa::Matrix &

—

inAlpha

oa::F32

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::leakyReluBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput,
	oa::F32 inAlpha = 0.01f
)

leakyReluBwd: backward pass for LeakyReLU activation. Computes: dInput = dOutput * (x > 0 ? 1 : alpha)

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

inAlpha

oa::F32

Default: 0.01f

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linear(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias = Matrix{}
)

linear: linear layer (fully connected). output = input @ weight^T + bias

Parameters

inX

const Matrix &

[batch, inFeatures]

inWeight

const Matrix &

[outFeatures, inFeatures]

inBias

const Matrix &

[outFeatures] (optional)

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearDataBwd(
	const Matrix & inGradOutput,
	const Matrix & inWeight
)

linearDataBwd: backward for linear layer (input gradient). Computes: dInput = dOutput @ weight^T

Parameters

inGradOutput

const Matrix &

—

inWeight

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearDataReluBwd(
	const Matrix & inGradOutput,
	const Matrix & inWeight,
	const Matrix & inActivation
)

linearDataReluBwd: fused linear data gradient followed by ReLU backward. Computes: dInput = (dOutput @ weight) * (activation > 0)

Parameters

inGradOutput

const Matrix &

—

inWeight

const Matrix &

—

inActivation

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearGelu(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias
)

linearGelu: fused linear + GELU. output = GELU(input @ weight^T + bias) The fused forward discards the pre-activation; the autograd node recomputes it (one GEMM) for geluBwd in the backward pass.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearRelu(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias
)

linearRelu: fused linear + ReLU. output = reLU(input @ weight^T + bias)

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearReluBwdData(
	const Matrix & inGradOutput,
	const Matrix & inWeight,
	const Matrix & inActivation
)

linearReluBwdData: fused in-layer linearRelu(x,W,b) backward, data path. For y = reLU(x @ W^T + b), computes dx = (dy * (act > 0)) @ W in a single dispatch. gate is applied INSIDE the inner sum (no materialization of dz), the opposite fusion direction of linearDataReluBwd.

Parameters

inGradOutput

const Matrix &

—

inWeight

const Matrix &

—

inActivation

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::linearSilu(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias
)

linearSilu: fused linear + SiLU. output = siLU(input @ weight^T + bias)

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

Returns

Matrix

The declared return value.

LinearWeightBiasBwdResult oa::FnMatrix::linearWeightBiasBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

linearWeightBiasBwd: fused weight and bias gradient for linear layer. Computes: dWeight = input^T @ dOutput, dBias = sum(dOutput, dim=0)

Parameters

inInput

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

LinearWeightBiasBwdResult

The declared return value.

Matrix oa::FnMatrix::linearWeightBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

linearWeightBwd: backward for linear layer (weight gradient). Computes: dWeight = input^T @ dOutput

Parameters

inInput

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::log(
	const oa::Matrix & inA
)

log: Element-wise natural log: out = log(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::logSoftmax(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

logSoftmax: Stable log-softmax over the selected dimension; -1 selects the last dimension.

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::logSoftmaxBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput,
	oa::I32 inDim = -1
)

No public source comment is attached to this declaration.

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

inDim

oa::I32

Default: -1

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::mamba3Mimo(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const Matrix & inMimoX,
	const Matrix & inMimoZ,
	const Matrix & inMimoO,
	const Matrix & inNormWeight,
	const SsmConfig & inConfig
)

--- Mamba-3 MIMO selective scan ---

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inMimoX

const Matrix &

—

inMimoZ

const Matrix &

—

inMimoO

const Matrix &

—

inNormWeight

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

Mamba3MimoBwdResult oa::FnMatrix::mamba3MimoBwd(
	const Matrix & inDOut,
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const Matrix & inMimoX,
	const Matrix & inMimoZ,
	const Matrix & inMimoO,
	const Matrix & inNormWeight,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inMimoX

const Matrix &

—

inMimoZ

const Matrix &

—

inMimoO

const Matrix &

—

inNormWeight

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Mamba3MimoBwdResult

The declared return value.

Matrix oa::FnMatrix::mamba3MimoStep(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const Matrix & inMimoX,
	const Matrix & inMimoZ,
	const Matrix & inMimoO,
	const Matrix & inNormWeight,
	const Matrix & inSsmState,
	const Matrix & inAngleState,
	const Matrix & inKState,
	const Matrix & inVState,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inMimoX

const Matrix &

—

inMimoZ

const Matrix &

—

inMimoO

const Matrix &

—

inNormWeight

const Matrix &

—

inSsmState

const Matrix &

—

inAngleState

const Matrix &

—

inKState

const Matrix &

—

inVState

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

Mamba3PreprocessResult oa::FnMatrix::mamba3Preprocess(
	const Matrix & inProjected,
	const Matrix & inDtBias,
	const Mamba3PreprocessConfig & inConfig
)

--- Mamba3Preprocess: fused in_proj split + RMSNorm + dt + A·dt ---

Parameters

inProjected

const Matrix &

—

inDtBias

const Matrix &

—

inConfig

const Mamba3PreprocessConfig &

—

Returns

Mamba3PreprocessResult

The declared return value.

Mamba3PreprocessBwdResult oa::FnMatrix::mamba3PreprocessBwd(
	const Matrix & inProjected,
	const Matrix & inDtBias,
	const Matrix & inDZ,
	const Matrix & inDX,
	const Matrix & inDBh,
	const Matrix & inDCh,
	const Matrix & inDDT,
	const Matrix & inDADT,
	const Matrix & inDTrap,
	const Matrix & inDAngle,
	const Mamba3PreprocessConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inProjected

const Matrix &

—

inDtBias

const Matrix &

—

inDZ

const Matrix &

—

inDX

const Matrix &

—

inDBh

const Matrix &

—

inDCh

const Matrix &

—

inDDT

const Matrix &

—

inDADT

const Matrix &

—

inDTrap

const Matrix &

—

inDAngle

const Matrix &

—

inConfig

const Mamba3PreprocessConfig &

—

Returns

Mamba3PreprocessBwdResult

The declared return value.

Matrix oa::FnMatrix::mamba3Siso(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const SsmConfig & inConfig
)

--- Mamba-3 SISO selective scan ---

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

SsmBwdResult oa::FnMatrix::mamba3SisoBwd(
	const Matrix & inDOut,
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

SsmBwdResult

The declared return value.

Matrix oa::FnMatrix::mamba3SisoStep(
	const Matrix & inC,
	const Matrix & inB,
	const Matrix & inX,
	const Matrix & inZ,
	const Matrix & inAdt,
	const Matrix & inDt,
	const Matrix & inTrap,
	const Matrix & inAngle,
	const Matrix & inCBias,
	const Matrix & inBBias,
	const Matrix & inD,
	const Matrix & inSsmState,
	const Matrix & inAngleState,
	const Matrix & inKState,
	const Matrix & inVState,
	const SsmConfig & inConfig
)

No public source comment is attached to this declaration.

Parameters

inC

const Matrix &

—

inB

const Matrix &

—

inX

const Matrix &

—

inZ

const Matrix &

—

inAdt

const Matrix &

—

inDt

const Matrix &

—

inTrap

const Matrix &

—

inAngle

const Matrix &

—

inCBias

const Matrix &

—

inBBias

const Matrix &

—

inD

const Matrix &

—

inSsmState

const Matrix &

—

inAngleState

const Matrix &

—

inKState

const Matrix &

—

inVState

const Matrix &

—

inConfig

const SsmConfig &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::maskedCategoricalAccuracyCount(
	const Matrix & inLogits,
	const Matrix & inLabels,
	const Matrix & inMask
)

No public source comment is attached to this declaration.

Parameters

inLogits

const Matrix &

—

inLabels

const Matrix &

—

inMask

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::matMulNt(
	const oa::Matrix & inA,
	const oa::Matrix & inB,
	oa::MatMulPrecision inPrecision = oa::MatMulPrecision::Auto
)
oa::Matrix oa::FnMatrix::matMulNt(
	const oa::Matrix & inInput,
	const oa::QuantMatrix & inWeight
)

matMulNt: oa::Matrix multiplication: B is [N,K] and out = A @ Bᵀ (the OA weight convention, shared with Linear/attention; NOT PyTorch-standard A with B as [K,N]). For batched/standard A use Bmm. Router resolves precision at context execution.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

inPrecision

oa::MatMulPrecision

Default: oa::MatMulPrecision::Auto

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::max(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

max: Reduce all values by maximum; dim is retained as a compatibility parameter and is not a dimensional reduction.

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::maxBwd(
	const Matrix & inInput,
	const Matrix & inMaxValue,
	const Matrix & inGradOutput
)

No public source comment is attached to this declaration.

Parameters

inInput

const Matrix &

—

inMaxValue

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

MaxPool2dResult oa::FnMatrix::maxPool2d(
	const Matrix & inX,
	oa::I32 inKernelSize,
	oa::I32 inStride,
	oa::I32 inPadding
)

No public source comment is attached to this declaration.

Parameters

inX

const Matrix &

—

inKernelSize

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

Returns

MaxPool2dResult

The declared return value.

Matrix oa::FnMatrix::maxPool2dBwd(
	const Matrix & inX,
	const Matrix & inIndices,
	const Matrix & inGradOutput,
	oa::I32 inKernelSize,
	oa::I32 inStride,
	oa::I32 inPadding
)

maxPool2dBwd: backward for 2D max pooling.

Parameters

inX

const Matrix &

—

inIndices

const Matrix &

—

inGradOutput

const Matrix &

—

inKernelSize

oa::I32

—

inStride

oa::I32

—

inPadding

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::mean(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

mean: Reduce by arithmetic mean; a valid non-negative Dim produces a keep-dimension axis reduction, while -1 produces a full reduction.

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::mergeHeads(
	const Matrix & inX,
	oa::I32 inBatch,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads
)

mergeHeads: [B*H,S,D/H] -> [B*S,D]. exact inverse of splitHeads.

Parameters

inX

const Matrix &

—

inBatch

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::mish(
	const oa::Matrix & inA
)

mish: Mish activation: out = A * tanh(softplus(A)).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::mishBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

mishBwd: backward pass for Mish activation.

Parameters

inInput

const Matrix &

forward INPUT x (mish'(x) depends on x directly)

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::moeCombine(
	const Matrix & inPacked,
	const Matrix & inRouteGate,
	const Matrix & inInverse,
	const Matrix & inPackedSlot
)

No public source comment is attached to this declaration.

Parameters

inPacked

const Matrix &

—

inRouteGate

const Matrix &

—

inInverse

const Matrix &

—

inPackedSlot

const Matrix &

—

Returns

Matrix

The declared return value.

MoeCombineBwdResult oa::FnMatrix::moeCombineBwd(
	const Matrix & inDOut,
	const Matrix & inPacked,
	const Matrix & inRouteGate,
	const Matrix & inInverse,
	const Matrix & inPackedSlot
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inPacked

const Matrix &

—

inRouteGate

const Matrix &

—

inInverse

const Matrix &

—

inPackedSlot

const Matrix &

—

Returns

MoeCombineBwdResult

The declared return value.

MoeExpertPlan oa::FnMatrix::moeExpertPlan(
	const Matrix & inExpertIndices,
	oa::I32 inNumExperts
)

No public source comment is attached to this declaration.

Parameters

inExpertIndices

const Matrix &

—

inNumExperts

oa::I32

—

Returns

MoeExpertPlan

The declared return value.

Matrix oa::FnMatrix::moeGather(
	const Matrix & inSelf,
	const Matrix & inIndices,
	const Matrix & inInverse
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

inIndices

const Matrix &

—

inInverse

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::moeGatherBwd(
	const Matrix & inSource,
	const Matrix & inInverse,
	oa::I32 inOutRows
)

No public source comment is attached to this declaration.

Parameters

inSource

const Matrix &

—

inInverse

const Matrix &

—

inOutRows

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::moeRouteWeights(
	const Matrix & inProbs,
	const Matrix & inExpertIndices
)

No public source comment is attached to this declaration.

Parameters

inProbs

const Matrix &

—

inExpertIndices

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::moeRouteWeightsBwd(
	const Matrix & inDOut,
	const Matrix & inProbs,
	const Matrix & inExpertIndices,
	const Matrix & inRouteWeights
)

No public source comment is attached to this declaration.

Parameters

inDOut

const Matrix &

—

inProbs

const Matrix &

—

inExpertIndices

const Matrix &

—

inRouteWeights

const Matrix &

—

Returns

Matrix

The declared return value.

void oa::FnMatrix::moeRoutingBiasUpdate(
	const Matrix & inSelectionMask,
	Matrix & inOutBias,
	oa::I32 inExpertsPerToken,
	oa::F32 inGamma
)

No public source comment is attached to this declaration.

Parameters

inSelectionMask

const Matrix &

—

inOutBias

Matrix &

—

inExpertsPerToken

oa::I32

—

inGamma

oa::F32

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::mul(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

mul: Element-wise multiplication: out = A * B.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::mulInPlace(
	Matrix & inSelf,
	const Matrix & inOther
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inOther

const Matrix &

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::neg(
	const oa::Matrix & inA
)

neg: Element-wise negation: out = -A.

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::ones(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::philoxNormal(
	const oa::Matrix & inA,
	oa::F32 inMean = 0.0F,
	oa::F32 inStddev = 1.0F,
	oa::U64 inSeed = 0
)

philoxNormal: Generate normal-distribution random values with Philox.

Parameters

inA

const oa::Matrix &

—

inMean

oa::F32

Default: 0.0F

inStddev

oa::F32

Default: 1.0F

inSeed

oa::U64

Default: 0

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::philoxUniform(
	const oa::Matrix & inA,
	oa::F32 inLow = 0.0F,
	oa::F32 inHigh = 1.0F,
	oa::U64 inSeed = 0
)

philoxUniform: Generate uniform-distribution random values with Philox.

Parameters

inA

const oa::Matrix &

—

inLow

oa::F32

Default: 0.0F

inHigh

oa::F32

Default: 1.0F

inSeed

oa::U64

Default: 0

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::pow(
	const oa::Matrix & inA,
	oa::F32 inExponent
)

pow: Element-wise power: out = A^Exponent.

Parameters

inA

const oa::Matrix &

—

inExponent

oa::F32

—

Returns

oa::Matrix

The declared return value.

oa::QuantMatrix oa::FnMatrix::quantize(
	const oa::Matrix & inInput,
	oa::Quantization inQuantization
)

quantize Float32 storage into one native OA Q4 or Q8 weight value. The returned value retains the source shape; its physical planes remain private.

Parameters

inInput

const oa::Matrix &

—

inQuantization

oa::Quantization

—

Returns

oa::QuantMatrix

The declared return value.

Matrix oa::FnMatrix::rand(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::randGlorotUniform(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::randKaimingUniform(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::randN(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::randXavier(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::reciprocal(
	const oa::Matrix & inA
)

reciprocal: Element-wise reciprocal: out = 1 / A.

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::relu(
	const oa::Matrix & inA
)

relu: ReLU activation: out = max(0, A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::reluBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput
)

reluBwd: backward pass for ReLU activation. Computes: dInput = dOutput * (forwardOutput > 0)

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::repeatInterleave(
	const Matrix & inA,
	oa::I32 inRepeats,
	oa::I32 inDim
)

Shape helpers

Parameters

inA

const Matrix &

—

inRepeats

oa::I32

—

inDim

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::repeatInterleaveBwd(
	const Matrix & inGradOut,
	MatrixShape inInputShape,
	oa::I32 inRepeats,
	oa::I32 inDim
)

No public source comment is attached to this declaration.

Parameters

inGradOut

const Matrix &

—

inInputShape

MatrixShape

—

inRepeats

oa::I32

—

inDim

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::reshape(
	const oa::Matrix & inA,
	oa::MatrixShape inShape
)

reshape: Reshape tensor to new shape (view operation, no copy).

Parameters

inA

const oa::Matrix &

—

inShape

oa::MatrixShape

—

Returns

oa::Matrix

The declared return value.

ResidualRmsNormResult oa::FnMatrix::residualRmsNorm(
	const Matrix & inA,
	const Matrix & inB,
	const Matrix & inWeight,
	oa::F32 inEps
)

--- ResidualRmsNorm: fused residual + RmsNorm ---

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

inWeight

const Matrix &

—

inEps

oa::F32

—

Returns

ResidualRmsNormResult

The declared return value.

oa::Matrix oa::FnMatrix::rmsNorm(
	const oa::Matrix & inSelf,
	const oa::Matrix & inWeight,
	oa::F32 inEps
)

rmsNorm: Weighted root mean square normalization over the last dimension: out = (A / sqrt(mean(A^2) + Eps)) * Weight.

Parameters

inSelf

const oa::Matrix &

—

inWeight

const oa::Matrix &

—

inEps

oa::F32

—

Returns

oa::Matrix

The declared return value.

RmsNormBwdResult oa::FnMatrix::rmsNormBwd(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inGradOutput,
	oa::F32 inEps = 1e-5F
)

rmsNormBwd: backward pass for RmsNorm.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inGradOutput

const Matrix &

—

inEps

oa::F32

Default: 1e-5F

Returns

RmsNormBwdResult

The declared return value.

Matrix oa::FnMatrix::rmsNormGated(
	const Matrix & inSelf,
	const Matrix & inWeight,
	const Matrix & inBias,
	const Matrix & inZ,
	oa::F32 inEps,
	bool inNormBeforeGate = true
)

LayerNorm and RmsNorm forward declarations are generated from MlFnMatrixNorm.toml.

Parameters

inSelf

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inZ

const Matrix &

—

inEps

oa::F32

—

inNormBeforeGate

bool

Default: true

Returns

Matrix

The declared return value.

RmsNormGatedBwdResult oa::FnMatrix::rmsNormGatedBwd(
	const Matrix & inX,
	const Matrix & inWeight,
	const Matrix & inBias,
	const Matrix & inZ,
	const Matrix & inGradOutput,
	oa::F32 inEps
)

rmsNormGatedBwd: backward for rmsNormGated (normBeforeGate = true). Returns grads w.r.t. x, weight, bias, z.

Parameters

inX

const Matrix &

—

inWeight

const Matrix &

—

inBias

const Matrix &

—

inZ

const Matrix &

—

inGradOutput

const Matrix &

—

inEps

oa::F32

—

Returns

RmsNormGatedBwdResult

The declared return value.

Matrix oa::FnMatrix::rnnCellLinear(
	const Matrix & inGatesI,
	const Matrix & inHidden,
	const Matrix & inWeightHh,
	oa::U32 inTimeOffset = 0,
	oa::U32 inBatchStride = 1,
	const Matrix & inBiasHh = Matrix{}
)

rnnCellLinear: fused vanilla-RNN recurrent step — Linear(h, W_hh) + rnnCellPointwise. Replaces the per-timestep pair of dispatches with one kernel. inGatesI is the whole input projection [B*T, H]; inTimeOffset/inBatchStride index the current timestep's row directly, so oa::Rnn needs no per-step Slice. The hidden projection required by reverse mode is retained internally rather than exposed as an output parameter.

Parameters

inGatesI

const Matrix &

—

inHidden

const Matrix &

—

inWeightHh

const Matrix &

—

inTimeOffset

oa::U32

Default: 0

inBatchStride

oa::U32

Default: 1

inBiasHh

const Matrix &

Default: Matrix{}

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::rnnCellPointwise(
	const Matrix & inGatesI,
	const Matrix & inGatesH
)

rnnCellPointwise: fused vanilla-RNN pointwise forward, h_new = tanh(gatesI + gatesH).

Parameters

inGatesI

const Matrix &

[B, H] input projection (W_ih x + b_ih)

inGatesH

const Matrix &

[B, H] hidden projection (W_hh h_prev + b_hh)

Returns

Matrix

The declared return value.

RnnCellPointwiseBwdResult oa::FnMatrix::rnnCellPointwiseBwd(
	const Matrix & inGatesI,
	const Matrix & inGatesH,
	const Matrix & inGradOutput,
	oa::I32 inHiddenSize,
	oa::U32 inTimeOffset = 0,
	oa::U32 inBatchStride = 1
)

rnnCellPointwiseBwd: fused vanilla-RNN pointwise backward. Returns gradients w.r.t. gatesI and gatesH (both equal to dL/da). inGatesI is the whole-sequence [B*T, H] projection; inTimeOffset/inBatchStride select this timestep's rows so dGatesI is scattered into the full buffer (zeros elsewhere).

Parameters

inGatesI

const Matrix &

—

inGatesH

const Matrix &

—

inGradOutput

const Matrix &

—

inHiddenSize

oa::I32

—

inTimeOffset

oa::U32

Default: 0

inBatchStride

oa::U32

Default: 1

Returns

RnnCellPointwiseBwdResult

The declared return value.

RnnScanResult oa::FnMatrix::rnnScan(
	const Matrix & inGatesI,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::I32 inSeqLen,
	oa::I32 inBatch,
	const Matrix & inBiasHh = Matrix{}
)

rnnScan: whole-sequence vanilla-RNN recurrent scan in ONE dispatch. Mathematically identical to running rnnCellLinear for each timestep, but collapses S dispatches into 1. The recurrent weight/bias gradient is computed separately via linearWeightBiasBwd on the saved hPrev.

Parameters

inGatesI

const Matrix &

[B*S, H] precomputed input projection (row b*S+t = timestep t)

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inSeqLen

oa::I32

—

inBatch

oa::I32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

RnnScanResult

The declared return value.

RnnScanBwdResult oa::FnMatrix::rnnScanBwd(
	const Matrix & inDOut,
	const Matrix & inGatesI,
	const Matrix & inHPrev,
	const Matrix & inWeightHh,
	oa::I32 inHiddenSize,
	oa::I32 inSeqLen,
	oa::I32 inBatch,
	const Matrix & inBiasHh = Matrix{}
)

rnnScanBwd: BPTT recurrence scan (backward of rnnScan) in ONE dispatch.

Parameters

inDOut

const Matrix &

—

inGatesI

const Matrix &

—

inHPrev

const Matrix &

—

inWeightHh

const Matrix &

—

inHiddenSize

oa::I32

—

inSeqLen

oa::I32

—

inBatch

oa::I32

—

inBiasHh

const Matrix &

Default: Matrix{}

Returns

RnnScanBwdResult

The declared return value.

Matrix oa::FnMatrix::sampleLogits(
	const Matrix & inLogits,
	oa::F32 inTemperature = 1.0F,
	oa::I32 inTopK = 0,
	oa::F32 inTopP = 1.0F,
	oa::U64 inSeed = 0
)

samples one class per last-axis row. temperature < = 0 is greedy; inTopK < = 0 keeps the full vocabulary; inTopP is clamped to (0,1].

Parameters

inLogits

const Matrix &

—

inTemperature

oa::F32

Default: 1.0F

inTopK

oa::I32

Default: 0

inTopP

oa::F32

Default: 1.0F

inSeed

oa::U64

Default: 0

Returns

Matrix

The declared return value.

oa::F32 oa::FnMatrix::scalar(
	const Matrix & inSrc
)

Extract first element as F32 scalar (requires single-element matrix).

Parameters

inSrc

const Matrix &

—

Returns

oa::F32

The declared return value.

oa::Matrix oa::FnMatrix::scale(
	const oa::Matrix & inA,
	oa::F32 inScalar
)

scale: Scalar multiplication: out = A * Scalar.

Parameters

inA

const oa::Matrix &

—

inScalar

oa::F32

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::scaleInPlace(
	Matrix & inSelf,
	oa::F32 inScalar
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inScalar

oa::F32

—

Returns

void

The declared return value.

Matrix oa::FnMatrix::scatterAddRows(
	const Matrix & inSource,
	const Matrix & inIndices,
	oa::I32 inOutRows
)

No public source comment is attached to this declaration.

Parameters

inSource

const Matrix &

—

inIndices

const Matrix &

—

inOutRows

oa::I32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::scatterRows(
	const Matrix & inSelf,
	const Matrix & inSource,
	const Matrix & inRowMap,
	const Matrix & inCount
)
Matrix oa::FnMatrix::scatterRows(
	const Matrix & inSelf,
	const Matrix & inSource,
	const CompactRowsResult & inPlan
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

inSource

const Matrix &

—

inRowMap

const Matrix &

—

inCount

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::scatterRowsBwdSource(
	const Matrix & inGradOut,
	const Matrix & inRowMap,
	const Matrix & inCount
)
Matrix oa::FnMatrix::scatterRowsBwdSource(
	const Matrix & inGradOut,
	const Matrix & inRowMap,
	const Matrix & inCount,
	const Matrix & inDispatchArgs
)

No public source comment is attached to this declaration.

Parameters

inGradOut

const Matrix &

—

inRowMap

const Matrix &

—

inCount

const Matrix &

—

Returns

Matrix

The declared return value.

void oa::FnMatrix::setRngSeed(
	oa::U64 inSeed
)

--- RNG --- seed the host-side seed generator. call once at startup for reproducibility.

Parameters

inSeed

oa::U64

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::sigmoid(
	const oa::Matrix & inA
)

sigmoid: Sigmoid activation: out = 1 / (1 + exp(-A)).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::sigmoidBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput
)

sigmoidBwd: backward pass for Sigmoid activation. Computes: dInput = dOutput * sigmoid(x) * (1 - sigmoid(x))

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::silu(
	const oa::Matrix & inA
)

silu: SiLU (Swish) activation: out = A * sigmoid(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::siluBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

siluBwd: backward pass for SiLU activation.

Parameters

inInput

const Matrix &

forward INPUT x (silu'(x) is a function of input, not output)

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::siluMul(
	const oa::Matrix & inA,
	oa::U32 inIntermediateSize
)

siluMul: SiLU on first half, multiply with second half: out = SiLU(A[:N]) * A[N:].

Parameters

inA

const oa::Matrix &

—

inIntermediateSize

oa::U32

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::siluMulBwd(
	const Matrix & inInput,
	const Matrix & inGradOutput
)

siluMulBwd: backward pass for SiluMul activation.

Parameters

inInput

const Matrix &

forward INPUT (gate||up); the forward output is not invertible

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::sin(
	const oa::Matrix & inA
)

sin: Element-wise sine: out = sin(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::slice(
	const oa::Matrix & inSelf,
	oa::I32 inDim,
	oa::I64 inStart,
	oa::I64 inEnd
)

slice: Materialize the half-open interval [Start, End) along one dimension.

Parameters

inSelf

const oa::Matrix &

—

inDim

oa::I32

—

inStart

oa::I64

—

inEnd

oa::I64

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::sliceBwd(
	MatrixShape inInputShape,
	oa::I32 inDim,
	oa::I64 inStart,
	oa::I64 inEnd,
	const Matrix & inDOut
)

No public source comment is attached to this declaration.

Parameters

inInputShape

MatrixShape

—

inDim

oa::I32

—

inStart

oa::I64

—

inEnd

oa::I64

—

inDOut

const Matrix &

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::softmax(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

softmax: Stable softmax over the selected dimension; -1 selects the last dimension.

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::softmaxBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput,
	oa::I32 inDim = -1
)

backward ops

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

inDim

oa::I32

Default: -1

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::softmaxScaledMasked(
	const Matrix & inScores,
	const Matrix & inMask,
	oa::F32 inScale
)

softmaxScaledMasked: fused transformer attention score normalisation. Computes: out = softmax(scores * scale + mask) over the last dimension.

Parameters

inScores

const Matrix &

—

inMask

const Matrix &

—

inScale

oa::F32

—

Returns

Matrix

The declared return value.

Matrix oa::FnMatrix::softmaxScaledMaskedBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput,
	oa::F32 inScale
)

softmaxScaledMaskedBwd: backward for the fused attention score op. Returns dScores = softmaxOut * (dOut - sum(dOut * softmaxOut)) * scale.

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

inScale

oa::F32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::softplus(
	const oa::Matrix & inA
)

softplus: Softplus: out = log(1+exp(A)), computed stably as max(A,0)+log1p(exp(-|A|)) (no f32 overflow).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::softplusBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput
)

softplusBwd: backward pass for Softplus activation.

Parameters

inForwardOutput

const Matrix &

output y = softplus(a) from the forward pass

inGradOutput

const Matrix &

—

Returns

Matrix

dOut * sigmoid(a) = dOut * (1 - e^-y)

oa::Vector<Matrix> oa::FnMatrix::split(
	const Matrix & inSelf,
	oa::Span<oa::I64> inSizes,
	oa::I32 inDim = 0
)

No public source comment is attached to this declaration.

Parameters

inSelf

const Matrix &

—

inSizes

oa::Span<oa::I64>

—

inDim

oa::I32

Default: 0

Returns

oa::Vector<Matrix>

The declared return value.

Matrix oa::FnMatrix::splitHeads(
	const Matrix & inX,
	oa::I32 inBatch,
	oa::I32 inSeqLen,
	oa::I32 inNumHeads
)

splitHeads: [B*S,D] -> [B*H,S,D/H]. Multi-head permutation.

Parameters

inX

const Matrix &

—

inBatch

oa::I32

—

inSeqLen

oa::I32

—

inNumHeads

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::sqrt(
	const oa::Matrix & inA
)

sqrt: Element-wise square root: out = sqrt(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::sub(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

Element-wise binary ops — broadcast-aware.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::subInPlace(
	Matrix & inSelf,
	const Matrix & inOther
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inOther

const Matrix &

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::subScalar(
	const oa::Matrix & inA,
	oa::F32 inScalar
)

subScalar: Scalar subtraction: out = A - Scalar.

Parameters

inA

const oa::Matrix &

—

inScalar

oa::F32

—

Returns

oa::Matrix

The declared return value.

void oa::FnMatrix::subScalarInPlace(
	Matrix & inSelf,
	oa::F32 inScalar
)

No public source comment is attached to this declaration.

Parameters

inSelf

Matrix &

—

inScalar

oa::F32

—

Returns

void

The declared return value.

oa::Matrix oa::FnMatrix::sum(
	const oa::Matrix & inA,
	oa::I32 inDim = -1
)

sum: Reduce by summation; a valid non-negative Dim produces a keep-dimension axis reduction, while -1 produces a full reduction.

Parameters

inA

const oa::Matrix &

—

inDim

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

oa::Matrix oa::FnMatrix::swiglu(
	const oa::Matrix & inA,
	const oa::Matrix & inB
)

swiglu: SwiGLU activation: out = SiLU(gate) * up.

Parameters

inA

const oa::Matrix &

—

inB

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

SwigluBwdResult oa::FnMatrix::swigluBwd(
	const Matrix & inGate,
	const Matrix & inUp,
	const Matrix & inOut,
	const Matrix & inGradOutput
)

swigluBwd: backward pass for SwiGLU activation.

Parameters

inGate

const Matrix &

—

inUp

const Matrix &

—

inOut

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

SwigluBwdResult

SwigluBwdResult with dGate, dUp

oa::Matrix oa::FnMatrix::tanh(
	const oa::Matrix & inA
)

tanh: Tanh activation: out = tanh(A).

Parameters

inA

const oa::Matrix &

—

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::tanhBwd(
	const Matrix & inForwardOutput,
	const Matrix & inGradOutput
)

tanhBwd: backward pass for Tanh activation. Computes: dInput = dOutput * (1 - tanh(x)^2), using saved tanh(x) output.

Parameters

inForwardOutput

const Matrix &

—

inGradOutput

const Matrix &

—

Returns

Matrix

The declared return value.

TopKResult oa::FnMatrix::topK(
	const Matrix & inA,
	oa::I32 inK,
	oa::I32 inDim = -1
)

No public source comment is attached to this declaration.

Parameters

inA

const Matrix &

—

inK

oa::I32

—

inDim

oa::I32

Default: -1

Returns

TopKResult

The declared return value.

Matrix oa::FnMatrix::topKMask(
	const Matrix & inIndices,
	oa::I32 inNumExperts
)

No public source comment is attached to this declaration.

Parameters

inIndices

const Matrix &

—

inNumExperts

oa::I32

—

Returns

Matrix

The declared return value.

oa::Matrix oa::FnMatrix::transpose(
	const oa::Matrix & inA,
	oa::I32 inDim0 = -2,
	oa::I32 inDim1 = -1
)

transpose: Materialize a swap of the last two axes for rank-2 or rank-3 matrices.

Parameters

inA

const oa::Matrix &

—

inDim0

oa::I32

Default: -2

inDim1

oa::I32

Default: -1

Returns

oa::Matrix

The declared return value.

Matrix oa::FnMatrix::upsampleBwd(
	const Matrix & inInput,
	const Matrix & inDOut,
	oa::I32 inScaleFactor,
	bool inIsBilinear
)

upsampleBwd: backward pass for Upsample (Nearest or Bilinear).

Parameters

inInput

const Matrix &

—

inDOut

const Matrix &

—

inScaleFactor

oa::I32

—

inIsBilinear

bool

—

Returns

Matrix

The declared return value.

VqAssignResult oa::FnMatrix::vqAssign(
	const Matrix & inZe,
	const Matrix & inCodebook
)

vqAssign: vector-quantization nearest-code assignment (VQ-VAE codebook lookup). inZe: [N, D] latents. inCodebook: [K, D] codes. Returns the per-row argmin index (int32 [N]) and the gathered winning code (float [N, D]) by squared L2 distance.

Parameters

inZe

const Matrix &

—

inCodebook

const Matrix &

—

Returns

VqAssignResult

The declared return value.

void oa::FnMatrix::vqEmaUpdate(
	const Matrix & inZe,
	const Matrix & inIdx,
	Matrix & ioEmbedSum,
	Matrix & ioClusterSize,
	Matrix & outCodebook,
	oa::F32 inDecay,
	oa::F32 inEps,
	oa::F32 inDeadThreshold,
	oa::U32 inSeed,
	bool inNormalize
)

vqEmaUpdate: EMA codebook update + dead-code reinit (van den Oord 2017). The codebook is NOT gradient-trained; each entry tracks the running mean of encoder outputs assigned to it, and dead codes are revived from live encoder rows. inZe [N,D], inIdx [N] int32 (from vqAssign), ioEmbedSum [K,D], ioClusterSize [K], outCodebook [K,D]. inDecay = EMA γ; inEps = division floor; inDeadThreshold = revive codes whose EMA count falls below it; inSeed = per-step seed; inNormalize rescales each codebook row to unit RMS (cosine VQ).

Parameters

inZe

const Matrix &

—

inIdx

const Matrix &

—

ioEmbedSum

Matrix &

—

ioClusterSize

Matrix &

—

outCodebook

Matrix &

—

inDecay

oa::F32

—

inEps

oa::F32

—

inDeadThreshold

oa::F32

—

inSeed

oa::U32

—

inNormalize

bool

—

Returns

void

The declared return value.

oa::ScalarType oa::FnMatrix::weightDtype()

--- Configuration --- weight allocation follows the precision of the active engine context.

Returns

oa::ScalarType

The declared return value.

Matrix oa::FnMatrix::zeros(
	MatrixShape inShape,
	oa::ScalarType inDtype = weightDtype()
)

No public source comment is attached to this declaration.

Parameters

inShape

MatrixShape

—

inDtype

oa::ScalarType

Default: weightDtype()

Returns

Matrix

The declared return value.